NegEx Algorithm

NegEx uses a curated lexicon of negation triggers (e.g., 'no', 'denies', 'without', 'absence of', 'free of') to flag sentences where clinical concepts may be negated. The algorithm scans text for these pseudo-negation indicators and applies directional rules to determine scope.

• Example: 'Patient denies chest pain' → 'chest pain' is tagged as negated
• Example: 'No evidence of pneumonia' → 'pneumonia' is tagged as negated
• Handles pre-negation (trigger before finding) and post-negation (trigger after finding)

NegEx operates on Unified Medical Language System (UMLS) concepts rather than raw strings, allowing it to recognize clinical findings regardless of surface form variation. The algorithm processes text that has already been mapped to standardized concept unique identifiers.

• Works with SNOMED CT, ICD-10-CM, and other UMLS source vocabularies
• Concept-level negation ensures 'no MI' and 'no myocardial infarction' are both correctly negated
• Enables downstream aggregation of negated findings across terminologies

NegEx constrains negation scope using a configurable token window between the trigger term and the target finding. If the clinical concept falls within the window, it is marked as negated; if it falls outside, the negation does not apply.

• Default window typically spans 5-6 tokens between trigger and finding
• Prevents over-negation when multiple findings appear in a single sentence
• Example: 'No chest pain, but patient reports shortness of breath' → only 'chest pain' is negated

NegEx includes a pseudo-negation filter that prevents false negation when trigger terms appear in non-negating contexts. Double negatives and hypothetical constructions are explicitly handled to avoid misclassification.

• 'Not only did the patient have fever' → 'fever' is not negated
• 'If symptoms do not improve' → hypothetical, not a factual negation
• 'Did not rule out infection' → double negative, finding remains affirmed
• Reduces false-positive negation rate in complex clinical narratives

NegEx achieves ~95% precision and ~80% recall on negation detection tasks without requiring GPU infrastructure or deep learning models. Its rule-based architecture makes it suitable for real-time clinical NLP pipelines.

• Processes thousands of documents per second on CPU
• Deterministic output ensures auditable and explainable results
• No training data required—works out of the box with UMLS-annotated text
• Often used as a baseline comparator when evaluating neural negation models

The ConText algorithm extends NegEx beyond simple negation to detect historical, hypothetical, and experiencer contexts. This allows systems to distinguish between current affirmed findings and those that occurred in the past or relate to someone other than the patient.

• 'Patient had pneumonia last year' → historical context, not currently active
• 'Mother has diabetes' → experiencer is family member, not patient
• 'Rule out myocardial infarction' → hypothetical/uncertain context
• Enables richer clinical temporal reasoning in medication reconciliation workflows

NegEx relies on the Unified Medical Language System (UMLS) to identify which terms in text represent clinical findings eligible for negation detection. The algorithm specifically targets UMLS semantic types such as:

• Disease or Syndrome (T047)
• Sign or Symptom (T184)
• Finding (T033)
• Laboratory or Test Result (T034) Without this ontological grounding, the algorithm cannot distinguish between a negatable finding and a procedure or anatomical reference.

The core mechanism of NegEx is a curated lexicon of negation phrases that invert the meaning of clinical findings within a defined window. Key trigger categories include:

• Pseudo-negation: "not ruled out" (double negative = affirmed)
• Pre-negation: "denies any chest pain" (trigger before finding)
• Post-negation: "cancer is ruled out" (trigger after finding) The original NegEx lexicon contains approximately 50 triggers, with modern implementations expanding to hundreds of domain-specific phrases.

NegEx applies a fixed token window (typically 5-6 tokens) between the negation trigger and the clinical finding to determine if negation applies. This window is directional:

• Forward scope: Trigger precedes finding ("no evidence of infiltrate")
• Backward scope: Trigger follows finding ("metastasis was ruled out") Termination rules stop the scope at sentence boundaries, conjunctions, or other clinical findings to prevent over-negation across clause boundaries.

Beyond binary negation, clinical NLP systems must detect hedging language that expresses diagnostic uncertainty. Common patterns include:

• Epistemic modifiers: "possible," "suspected," "cannot exclude"
• Evidential qualifiers: "appears consistent with," "suggestive of"
• Probabilistic framing: "low suspicion for," "atypical for" Systems like UncertaintyEx extend the NegEx framework to classify findings as affirmed, negated, or uncertain — a tripartite distinction critical for accurate clinical data extraction.

Clinical findings must be classified by temporal status to distinguish active problems from resolved history. Negation-adjacent temporal triggers include:

• Historical: "history of MI," "prior to admission," "status post"
• Recent/Resolved: "resolved pneumonia," "no longer present"
• Current: Implicitly assumed unless modified Failure to detect temporal context leads to problem list pollution, where resolved conditions persist in active diagnoses, degrading clinical decision support accuracy.